WayCode

An AI coding agent that lives in your VS Code sidebar — talks to you in your language, works on your whole project, and never changes a file behind your back.

What is WayCode?

WayCode is a VS Code extension that turns a chat panel into a real coding agent. You describe what you want — in English, Hebrew, Arabic, or anything else — and it reads your repository, searches the code, writes and edits files, runs your tests and linter, reads the errors, and fixes them. Every risky step shows you a diff or the exact command and waits for your approval, unless you tell it not to.

It is provider-agnostic: Claude, GPT, a local Ollama model, or your existing Claude Code CLI subscription. And it can split the work between two models — one that's great at your language and one that's great at code.

┌──────────────────────────────────────────────────────────────┐
│  You: "תוסיף בדיקות ל-History ותריץ אותן"                    │
├──────────────────────────────────────────────────────────────┤
│  🔍 search_code  /class History/            → 3 matches      │
│  📖 read_file    src/memory/History.ts      → 62 lines       │
│  ✅ create_file  src/test/history.test.ts   → [diff] approve?│
│  🧪 run_tests    npm test                   → 14 pass        │
├──────────────────────────────────────────────────────────────┤
│  WayCode: הוספתי 5 בדיקות ל-History. כל 14 הבדיקות עוברות.   │
└──────────────────────────────────────────────────────────────┘

Highlights

Quick start

1. Install — from the Marketplace, or build a .vsix locally (see Development).
2. Open the panel — click the WayCode icon in the Activity Bar, or run WayCode: Open Chat.
3. Pick a model — WayCode: Select Model / Provider.
4. Add a key if needed — WayCode: Set API Key (stored in VS Code Secret Storage, never in settings).
5. Ask for something — "add input validation to the login form and run the tests".

Two ways to run WayCode without paying per token:

• Local Ollama — install Ollama, pull a model (ollama pull qwen2.5-coder, ollama pull gemma2), then choose provider Ollama (local). The model picker lists what you have installed. Everything stays on your machine.
• Claude Code CLI — if you already have a Claude subscription with the claude CLI installed, choose provider Claude Code (CLI, no key). Note: the CLI provider is text-only, so use it as the communicator or in single-agent chat — it cannot drive WayCode's tools as the coder.

How it works

Single-agent mode (default)

One model does everything, in a bounded tool-use loop:

   your message
        │
        ▼
  ┌───────────┐   tool call    ┌────────────────────────┐
  │   model   │ ─────────────▶ │ read · search · edit    │
  │  (agent)  │ ◀───────────── │ terminal · tests · lint │
  └───────────┘    result      └────────────────────────┘
        │            ▲                    │
        │            └── approval ────────┘  (diff / command shown to you)
        ▼
   final answer            up to `maxAgentSteps` iterations (default 25)

If the model announces a plan but calls no tool, WayCode detects the stall and nudges it to act — so you don't get a to-do list instead of working code.

Multi-agent mode (waycode.multiAgent.enabled)

Two models, each doing what it's best at:

  you (any language)
        │
        ▼
  🗣️  COMMUNICATOR ──── CHAT? ────▶ answers you directly (no coder round-trip)
        │
      CODE? → precise English task spec
        │
        ▼
  👨‍💻  CODER  ──▶ tools ──▶ verify (tests / lint) ──▶ fix ──▶ repeat
        │
        ▼  ground truth: the exact tool calls that really ran
  🗣️  COMMUNICATOR ──▶ explains the result in your language

The communicator can be a small local model (default: gemma2 on Ollama) while the coder is a code-strong model — and it only ever reports what actually happened.

Modes

Switch with the mode button in the chat, or cycle with Shift+Tab.

Even in 🌙 Auto, an overwrite that would wipe most of an existing file still asks — a "summary" write that collapses a 500-line file is never approved silently.

Tools

Guardrails: every path is resolved inside the workspace (no ../ escapes), and commands that never exit (npm run dev, vite, tsc --watch, docker compose up, …) are refused with a pointer to a one-shot check instead — so a run can't burn its budget on a watcher. Tool names models commonly guess (bash, str_replace, ls, grep, …) are aliased to the real ones.

Providers

All providers share one HTTP layer with timeouts and readable errors, so a loading local model can't hang the panel forever.

Commands

Settings

Project memory

WayCode carries four kinds of memory into every request:

• WAYCODE.md — a per-project context file you own (also read from .waycode.md or .waycode/context.md). Put the stack, conventions, and build/test commands there; it's injected into every system prompt. Run WayCode: Edit Project Context File to scaffold it.
• Conversation context file — one per chat, written automatically. Each turn appends what you asked, the tools that actually ran, and the answer you were given. Both the communicator and the coder read it at the start of every turn, so a conversation keeps its context across reloads, history trimming, and reopening it from 🕘. Run WayCode: Open Conversation Context File to read it.
• Workspace memory — pinned files, recorded decisions, and preferences, persisted per workspace.
• Conversation history — past threads, saved globally and reopenable from 🕘. Loading one replays it into the model, not just onto the screen.

Development

npm install
npm run compile      # tsc -p ./
npm run watch        # rebuild on change
npm run typecheck    # tsc --noEmit
npm test             # node --test out/test/  (compiles first)
npm run lint
npx @vscode/vsce package   # → waycode-<version>.vsix

Press F5 in VS Code to launch an Extension Development Host with WayCode loaded. CI type-checks, compiles, and tests on Node 18 and 20.

Layout

src/
├─ extension.ts          activation + all commands
├─ config.ts             settings & Secret Storage
├─ agent/
│  ├─ Agent.ts           the plan → act → verify → fix loop
│  ├─ Orchestrator.ts    communicator ⇄ coder pipeline
│  ├─ prompts.ts         system prompts
│  ├─ routing.ts         CHAT vs CODE classification
│  └─ history.ts         context-window trimming
├─ providers/            Anthropic · OpenAI · Ollama · Claude CLI (one interface)
├─ tools/                file · search · command tools, diffs, path safety
├─ context/              project snapshot + WAYCODE.md
├─ memory/               workspace memory · conversation history · per-chat context file
├─ ui/                   chat webview + settings panel
└─ test/                 node:test suite

Adding a provider means implementing AIProvider and registering it in ProviderFactory; adding a tool means implementing Tool and listing it in ToolRegistry.default(). Nothing else changes.